Autograd for off-resonance support

[paquiteau]

• Closes None

Changes proposed in this pull request:

• Add Support for autograd wrapping off-resonance operators
• Fixed bug for array conversion with dict arguments.

[chaithyagr]

The PR seems to have very less changes than i assumed? Do things work straight Outta the box?

Is the autodiff of library approximation actually close to autodiff of complete off resonance operation? (Note how autodiff of NUFFT was not matching autodiff of ndft perfectly).

Or is the PR still not ready for review?

Tests and example can help!
A great example is trying to learn off resonance map from a decent initialisation with parameterization of maps by splines. WDYT?

[paquiteau]

Yes it's that simple to make it happen :)

Adding tests could be done, but I am also working/testing it in downstream application. However learning the field-map directly is a badly conditioned problem, and it also depends on the image/data itself in a intricated manner. I don't think we want to support this directly in mri-nufft (at least for now)